Kernel cross-view collaborative representation based classification for person re-identification

Raphael C. Prates, William Robson Schwartz

Introduction

Person Re-Identification (Re-ID) is a critical task in surveillance, involving the recognition of individuals across non-overlapping camera networks. This task is inherently challenging due to significant variations in lighting, camera viewpoints, and human poses. This paper addresses two primary limitations in current Re-ID systems: the Small-Sample-Size (SSS) problem, where limited training data hinders deep learning performance, and the Camera Transition Gap, which involves the complex, nonlinear relationship between images captured by different cameras.

Methodology: The Kernel X-CRC Framework

The authors propose Kernel Cross-View Collaborative Representation based Classification (Kernel X-CRC) to overcome these hurdles. The framework introduces three core innovations:

  1. Multi-Task Learning and Similarity Constraints: Unlike standard Collaborative Representation-based Classification (CRC) that treats probe and gallery images independently, Kernel X-CRC solves them simultaneously. It introduces a similarity term to the optimization objective, forcing the model to find consistent "coding vectors" for the same individual across different views. This encodes the prior knowledge that a person's identity remains constant despite camera changes.
  2. The Kernel Trick: To model the nonlinear appearance changes between cameras, the authors project features into a high-dimensional kernel space (using RBF or exponential $\chi^2$ kernels). This allows the model to capture complex relationships that linear CRC would miss.
  3. Analytical Efficiency: A significant advantage of Kernel X-CRC is its closed-form analytical solution. This avoids the slow, iterative processes required by many sparse representation methods, making it computationally viable for real-time applications.

Experimental Setup and Datasets

The model was rigorously evaluated on four benchmark datasets:

  • PRID 450S: 450 image pairs with high viewpoint and lighting variance.
  • VIPeR: 632 pairs characterized by extreme viewpoint changes (often exceeding 90 degrees).
  • CUHK01: 971 individuals captured in a campus setting.
  • GRID: A challenging dataset from an underground station featuring poor illumination and a gallery full of "distractors."

Key Results and Discussion

Kernel X-CRC demonstrated state-of-the-art or competitive performance across all benchmarks, particularly excelling in scenarios with limited training data. On the PRID 450S dataset, it achieved a 68.8% Rank-1 accuracy, a record at the time of publication. On the difficult GRID dataset, it reached 26.6% Rank-1 accuracy, significantly outperforming existing methods.

Table 3: Comparison of Kernel X-CRC against baseline collaborative representation methods on the VIPeR dataset.

As shown in the table above, the incremental additions of the cross-view constraint (X-CRC) and the kernel trick (Kernel X-CRC) lead to substantial improvements over standard SRC and CRC baselines. On the VIPeR dataset, the method achieved a 51.6% Rank-1 accuracy when paired with the GoG feature descriptor.

Table 1: Performance of various feature descriptors using the proposed classification framework on the VIPeR dataset.

Conclusion

Kernel X-CRC effectively bridges the gap between different camera views by balancing reconstruction accuracy with cross-view consistency. By leveraging the kernel trick and a similarity-constrained multi-task framework, it provides a robust solution to the SSS problem in person Re-ID. Future research may focus on integrating local region constraints to better handle heavy occlusions and scaling the kernel matrix computation for massive datasets.

Summary updated

Abstract

Person re-identification aims at the maintenance of a global identity as a person moves among non-overlapping surveillance cameras. It is a hard task due to different illumination conditions, viewpoints and the small number of annotated individuals from each pair of cameras (small-sample-size problem). Common subspace learning methods have been proposed to handle the camera transition problems. However, after learning the low-dimensional representation, these methods usually compute distances using a simple cosine or Mahalanobis distance. Therefore, an still open question is how to better match probe and gallery images in the learned common subspace considering reduced number of training samples and the nonlinear behavior of the data. Collaborative Representation based Classification (CRC) has been employed successfully to address the small-sample-size problem in computer vision. However, the original CRC formulation is not well-suited for person re-identification since it does not consider that probe and gallery samples are from different cameras. Furthermore, it is a linear model, while appearance changes caused by different camera conditions indicate a strong nonlinear transition between cameras. To overcome such limitations, we propose the Kernel Cross-View Collaborative Representation based Classification (Kernel X-CRC), method that represents probe and gallery images by balancing representativeness and similarity nonlinearly. According to experimental results, we achieve state-of-the-art for rank-1 matching rates in three person re-identification datasets (CUHK03, PRID450S and GRID) and the second best results on VIPeR and CUHK01 datasets. Furthermore, we present outperforming results on Market-1501 dataset demonstrating that the Kernel X-CRC is suitable to a large-scale and multiple cameras scenario.
BibTeX
@article{prates2019kernel,
  author = {Raphael C. Prates and William Robson Schwartz},
  title = {{Kernel cross-view collaborative representation based classification for person re-identification}},
  journal = {J. Vis. Commun. Image Represent.},
  year = {2019},
  doi = {10.1016/j.jvcir.2018.12.003},
}
Figure